Bülent Uyaniker PhD, Abstract Co-Author: Nothing to Disclose

Rasika Rajapakshe PhD, Presenter: Nothing to Disclose

Paula Beth Gordon MD, Abstract Co-Author: Expert Advisory Committee, The Medipattern Corporation Stockholder, mdconversation Spouse, Medical Advisory Board, GlaxoSmithKline plc Spouse, Medical Advisory Board, Protox Therapeutics Inc Spouse, stockholder, OncoGenex Pharmaceuticals Inc Spouse, stockholder, Genyous Biomed International

Stuart Faulkner Silver MD, Abstract Co-Author: Nothing to Disclose

To develop an automated method to estimate the percent breast density from digital mammograms.

The algorithm for estimating the breast density for a given mammogram is based on critical density variation criteria and its detection within the image and includes determinant analysis complemented by morphological image processing as well as statistical quantification, histogram based analysis, segmentation and clustering techniques. This method automatically separates the breast boundary, fatty tissue and breast parenchyma and calculates the percentage dense tissue with respect to whole breast. Two sets of digital mammography image samples, consist of 36 and 51 images (R-CC), have been assessed for breast density. The images were also assessed by two experienced breast radiologists using well established manual thresholding software (Cumulus) for breast density. These results were compared with the new automatic evaluation method. A second estimate of the percent breast density by radiologists from first image set, after an arbitrarily chosen six month time interval, provided a measure for intra-observer variability.

The comparison, using Pearson’s product-moment correlation, for the intra-observer variability resulted in r1=0.89 and r2=0.87 for the first data set for two radiologists, respectively. Similarly, the inter-observer comparison for the two radiologists yielded r1=0.89 and r2=0.85 for the first and second data sets. The automatic evaluation algorithm results in comparison to radiologists assessment gave r1=0.85 and r2=0.90 for the first data set and r1=0.86 and r2=0.85 for the second set. The significance level for all the coefficients was P < 0.0001. The correlations between the radiologists and automatic estimates imply that the explained variances in each case are comparable with the agreement between and among the two radiologists.

The results of the automatic method presented here indicate that the algorithm is a quantitative and reproducible way of assessing percent breast density in digital mammograms, and is ready for clinical implementation.

Availability of a fully automated, reproducible method of estimating breast density will enable the wide spread use of this information for breast cancer risk assessment.

Uyaniker, B, Rajapakshe, R, Gordon, P, Silver, S, A Fully Automatic Method for Estimating Breast Density in Digital Mammograms.  Radiological Society of North America 2009 Scientific Assembly and Annual Meeting, November 29 - December 4, 2009 ,Chicago IL. http://archive.rsna.org/2009/8013231.html